AI-Powered NLP Framework for Extracting Drug Safety Information in Pregnancy

Pregnancy presents a unique clinical scenario where the safety of pharmacological interventions is of paramount importance. The potential teratogenic risks associated with drug intake during pregnancy necessitate a highly informed, evidence-based approach to prescribing. However, the rapid increase in published literature, clinical trials, and drug safety communications poses significant challenges for clinicians attempting to stay current with evolving data. To address this, we propose an AI-powered Natural Language Processing (NLP) framework designed to extract, interpret, and classify drug safety information from unstructured, pregnancy-related clinical texts. The system leverages transformer-based deep learning models—specifically fine-tuned variants of BERT—to classify drug risk levels into five categories: Safe, Low, Medium, High, and Unknown. Importantly, our approach integrates trimester-specific analysis to account for temporal variability in drug risk, a factor often overlooked in generalized models. The framework includes a visual analytics dashboard that supports confidence scoring, risk visualization, and interactive querying by trimester or drug class. Our model was trained and validated using annotated data derived from FDA labels, CANMAT guidelines, WHO publications, and peer-reviewed case reports. Validation results demonstrate high precision in identifying clear-cut cases (e.g., Paracetamol as Safe; Warfarin as High risk) and meaningful generalization across unseen data. A functional prototype has been developed that supports clinicians in making real-time, literature-informed prescribing decisions. Ultimately, this research contributes to safer pharmacotherapy in pregnancy by providing a scalable, explainable, and clinically relevant AI tool. It serves as both a foundation for future research and a practical application for healthcare decision support.